Accurate identification of geological formations is essential for understanding tectonic structures, planning mining activities, and sustainably managing natural resources. It goes beyond the scientific framework to play a key role in economic development, environmental preservation, and population security. This article proposes a study using machine learning to analyze different parameters from various sources of satellite imagery: multispectral optics (Landsat-8), radar (ALOS PALSAR), and soil and morphometric parameters (soil, altitude, slope, curvature, and shady). The data were preprocessed to remove atmospheric biases and harmonize spatial resolutions. Techniques such as principal component analysis, band ratios, and image fusion have made it possible to enrich imagery by highlighting spectral and textural characteristics. Finally, classifiers such as Random Forest, Gradient Boosting, and XGBoost (version 1.6.2) were used to evaluate the impact of each parameter on the classification. The results show that geographic parameters combined with PCA provide the best overall performance with Random Forest, achieving an accuracy of 55.29% and an MCC of 45.12% while ensuring a rapid training speed (3.6 s). The geographic parameters associated with the OLI spectrometric data show a good balance, with XGBoost achieving a slightly higher MCC (40.3%) with a moderate training time (7.9 s). On the other hand, the OLI spectrometric parameters coupled with PCA display significantly lower performance, with an accuracy of 45.05% and an MCC of 31.81% for Random Forest. These observations highlight the potential of geographic and geological parameters associated with suitable models to improve classification. The multi-source approach thus proves optimal for more robust and precise results.
Read full abstract